sr-ai 2.4.1

AI backends, caching, and AI-powered git commands for sr
Documentation

sr-ai

AI backends, caching, and AI-powered git commands for sr.

crates.io

Overview

sr-ai provides the AI layer for sr. It includes:

  • AI backends — Claude, GitHub Copilot, and Gemini with automatic detection and fallback
  • Commands — commit, rebase, review, explain, branch, pr, ask, cache
  • Caching — fingerprint-based commit plan caching with incremental re-analysis

Safety & Sandboxing

All AI backends are sandboxed to prevent the agent from modifying the repository:

  • Read-only git — agents can only run diff, log, show, status, ls-files, rev-parse, branch, cat-file, rev-list, shortlog, blame. Mutating commands are blocked.
  • No shell access — agents cannot run arbitrary shell commands or delete files.
  • Working tree snapshots — sr commit snapshots the full working tree before invoking the agent. On failure, the snapshot is automatically restored. On success, it is cleared.
  • Programmatic mutations — all git writes (staging, committing) happen in sr's Rust code after the agent returns, never inside the agent.

Snapshots are stored in the platform data directory (~/.local/share/sr/snapshots/<repo-id>/ on Linux, ~/Library/Application Support/sr/snapshots/<repo-id>/ on macOS), keyed by a SHA-256 hash of the repository root path.

AI Backends

Backend CLI required Env var Default model
Claude claude — haiku
Copilot gh copilot — gpt-4.1
Gemini gemini — (default)

Backends are auto-detected in order: Claude, Copilot, Gemini. Use --backend or SR_BACKEND to override.

Commands

Command Description
commit Analyze changes and generate atomic conventional commits
rebase AI-powered interactive rebase (reword, squash, reorder commits)
review AI code review with severity-based feedback
explain Explain what a commit does and why
branch Suggest a conventional branch name
pr Generate PR title and body from branch commits
ask Freeform Q&A about the repository
cache Manage commit plan cache (status, clear)

Caching

Commit plans are cached at ~/.cache/sr/ai/<repo-id>/entries/. Cache features:

  • Exact hit — identical fingerprint match, reuse plan directly
  • Incremental hit — partial match, AI re-analyzes only changed files
  • TTL — entries expire after 24 hours
  • LRU — max 20 entries per repo

Usage

This crate is used as a library by sr-cli. It is not intended to be used directly, but can be embedded in other tools:

[dependencies]
sr-ai = "2"

Prerequisites

At least one AI backend CLI must be installed:

License

Apache-2.0